MIST: A Benchmark and Baseline for Multi-Frame Infrared Small Target Detection in Complex Motion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Multi-frame infrared small target detection (MISTD) is crucial, but current methods struggle with complex target motion.
• Existing datasets often feature simplified motion patterns, leading to data-driven models that fail in real-world scenarios with irregular or fast-moving targets.
• This results in noisy feature representations and poor detection performance.

Purpose of the Study:

• To address the limitations of existing MISTD datasets and methods by introducing a new large-scale dataset and a robust detection model.
• To enable research on MISTD algorithms capable of handling complex and diverse target motion patterns.
• To establish a challenging benchmark for evaluating MISTD algorithms in realistic airborne infrared detection scenarios.

Main Methods:

• Proposed the MIST dataset, a large-scale, synthetic dataset featuring low signal-to-clutter ratios and complex target motions in realistic backgrounds.
• Developed MISTNet, a novel baseline model employing the Information Bottleneck theory for robust feature representation.
• Introduced a shifted neighborhood compensation block for implicit motion compensation and a progressive distillation decoder for hierarchical information filtering.

Main Results:

• Evaluated 31 state-of-the-art MISTD methods on the MIST dataset, revealing significant performance drops compared to existing benchmarks.
• MISTNet demonstrated superior performance, outperforming all other methods by a substantial margin with over a 6% gain in the IoU metric.
• The results highlight the challenges posed by complex motion in MISTD and the effectiveness of the proposed MISTNet.

Conclusions:

• The MIST dataset provides a more realistic and challenging benchmark for MISTD research, particularly for complex motion scenarios.
• MISTNet offers a robust and effective solution for MISTD, outperforming existing methods by accurately handling irregular and fast target movements.
• Further research in MISTD should focus on developing algorithms that can generalize to complex motion patterns found in real-world applications.